Model Distillation Arms Race: 80+ Papers Show China's AI Military Shortcut

The AI community has long celebrated distillation as a way to compress models—but now we see its weaponization at scale. Chinese military labs are using teacher-student techniques to suck core competencies out of U.S. frontier models, turning multi-billion-dollar R&D into a free lunch. This isn't just about export controls on hardware; it's a fundamental challenge to the economic model of commercial AI. If distillation can strip a model's value at negligible cost, tech companies and governments must rethink how they release and protect their algorithms.

Chinese military researchers have systematically used outputs from leading U.S. AI models developed by OpenAI and Anthropic to train domestic defense systems, according to a Reuters review of more than 80 academic papers and patents. The findings, compiled with assistance from the Washington-based Jamestown Foundation, reveal a sophisticated effort to leverage model distillation—a technique where a smaller, specialized model learns from a larger, more capable one—to accelerate the development of AI for military applications while bypassing the massive computational requirements normally needed to train frontier systems from scratch.

“

AI models developed by OpenAI and Anthropic to train domestic defense systems, according to a Reuters review of more than 80 academic papers and patents.

This discovery lands at a critical moment in the U.S.-China technology rivalry. Washington has imposed stringent export controls on advanced semiconductors and chip-making equipment, aiming to limit Beijing's ability to produce high-performance AI processors. The use of model distillation, however, provides a workaround: Chinese researchers can tap the outputs of existing powerful U.S. models—accessible via cloud services or APIs—to train purpose-built AI systems that require only modest hardware. The papers reviewed by Reuters indicate that institutions linked to the People's Liberation Army (PLA) are among the most active adopters, applying distilled models to tasks ranging from imagery analysis and target recognition to command-and-control simulations.

The practice itself is not new; distillation is a legitimate and widely used technique in the AI industry. Companies routinely distill large models into smaller ones for cost-efficient deployment. But the dispute here centers on unauthorized extraction of capabilities from proprietary U.S. models, which American officials argue represents a theft of intellectual property and a direct circumvention of export controls. A senior U.S. official was quoted saying such practices "undermine our national security and the integrity of our technological leadership." China has responded by accusing Washington of pursuing AI "hegemonism" and insisting that U.S. companies have engaged in similar distillation of foreign models.

The timing intensifies the geopolitical stakes. The revelations emerge just ahead of planned U.S.-China talks on AI governance and safety. For the U.S., these talks will likely become a venue to demand stricter enforcement against unauthorized distillation, possibly through new restrictions on access to American AI services. For China, the discovery that its military is so deeply reliant on U.S. models could undercut its narrative of indigenous AI prowess, even as state-backed AI firms like Moonshot claim independent breakthroughs.

From a technical perspective, the 80-plus documents paint a picture of an organized, institutionalized effort. Researchers used outputs from OpenAI's GPT series and Anthropic's Claude family to train domain-specific models. These distilled models can then be deployed on edge devices or in classified environments without internet connectivity, making them harder to detect and disrupt. The PLA appears to have integrated this strategy into a broader "intelligentization" doctrine that prioritizes AI as a force multiplier.